| In recent years,the application of video surveillance systems in public areas or private places has become more and more widespread.And large-scale surveillance data has been generated.Since the timely detection of abnormalities in the surveillance video can provide clues for the security department to handle the case.It is of great significance to social and public security.Manual inspection is not only at particular risk of omitting but also time-consuming and labor-intensive.Therefore,video anomaly detection becomes an important part of surveillance systems.The anomaly data set is costly to obtain.In addition,the anomaly definition is ill-defined.In view of this,unsupervised video anomaly detection has become the main research direction of video anomaly detection.This thesis focuses on two unsupervised video anomaly detection technology based on memory modules and prediction.The research methods and innovations are as follows:(1)An anomaly detection model combining multi-scale memory module and multi-scale structural similarity is proposed.In order to solve the problem that the dynamic prototype unit model based on memory unit does not make full use of multilevel features and does not consider the structural differences between abnormal and normal events when detecting video anomalies.The new model constructs a multiscale memory module,which uses memory units of different scale space to encode the features of the encoding layer,and concatenates the encoding results with the features of the decoding layer,which can not only preserve the shallow details of the network,but also promote the diversity of normal patterns.In order to constrain the learning of structural information in normal events,the multi-scale structure similarity index error and mean absolute error are combined as objective functions.The proposed method makes the event structure in the predicted video closer to the normal event,and improves the competency of its prediction for abnormal events in videos.(2)Aiming at the problems of inaccurate prediction of normal video and poor ability of learning normal features in video anomaly detection,this thesis proposes an anomaly detection model combining wavelet transform and Encoder-Decoder Attention.The model introduces multi-level discrete wavelet transform,and designs a module of Discrete Wavelet Transform Fusion.The module concatenates the subbands obtained by decomposing video frames,and feeds the result into depthwise separable convolution,and then fuses with the encoder features to compensate for the high-frequency details lost in the downsampling process.The model also constructs Encoder-Decoder Attention module.After performing Difference of Gaussian operation on the encoder feature map,the attention weights are obtained along the horizontal and vertical directions respectively.Then the encoder features are aggregated according to the weights.Finally the decoder features are associated to enhance the network’s learning of normal events. |
